In today’s world, scene recognition is a major challenge in various computer vision applications ranging from autonomous navigation to content-based image retrieval. Conventional methods often struggle to capture the complex patterns present in complex scenes, leading to limited recognition accuracy. Deep learning, and in particular CNN, has evolved as a significant technology to automatically detect and extract hierarchical features from raw pixel data, significantly improving scene recognition performance. However, challenges such as environmental variability and scene confusion remain. Therefore, to solve this problem, the authors proposed a new hybrid model based on CNN and deep learning for scene recognition that solves this challenge by learning reliable features. The model includes innovative strategies to mitigate common problems such as redundancy and data asymmetry, thereby improving performance in real-world scenarios. Through extensive testing and evaluation on the Indoor-67 dataset, the obtained results explain the efficacy and scalability of the proposed approach compared to exiting methods. This research contributes to the development of the field of scene recognition by providing a comprehensive framework that combines deep learning with special optimization techniques to achieve high recognition accuracy in various scene categories.